strict-validation-setup

The skill ships two distinct concerns split by temporal phase:

• Project-stable — strict-mode tooling config + AGENTS.md authoring (defer the AGENTS.md content to init; this skill only ensures it exists and references the per-task pattern).
• Task-ephemeral — .agent-tasks/<task-id>/GOALS.md per task plus failing-test scaffolding co-located with it.

These never mix. Task goals never go into AGENTS.md (would leak as project policy). Project invariants never go into per-task GOALS.md (would duplicate per task and drift).

Modality differentiation

Skill	Owns
strict-validation-setup	Strict-mode tooling configs + per-task GOALS.md convention (this file)
init	AGENTS.md authoring (project-stable) — defer to it for content
test-driven	TDD discipline (RED → GREEN → REFACTOR) — defer for test-writing methodology
type-driven	Refined-type / typestate specs — defer for type-system invariants
design-by-contract	Pre/post conditions, runtime contracts — defer for assertion patterns
validation-first	State-machine specs (typestate / FSM / actor) — defer for FSM modeling
tests-adversarial	Assumption-violation tests — defer for the adversarial test pattern
setup-pre-commit	Commit-hook installation — defer for hook tooling
setup-gitignore	Gitignore patterns — defer for ignore-file composition

Duplication with the existing skills was audited and accepted; bodies cite each rather than re-doing the work. When the surface narrows to a single concern above, defer.

Three parts

1. Strict-mode tooling bootstrap (project-stable)

Detect the ecosystem from manifests, then write strict-mode config per the relevant references/<ecosystem>.md. Idempotency: merge with existing config; raise if a destructive overwrite would be required and --overwrite is not explicit. Never silently replace.

Languages with bundled references (Q5-approved set; framework specifics deferred to a follow-up):

• references/typescript.md
• references/python.md
• references/rust.md
• references/go.md
• references/ocaml.md

Languages noted but not yet bundled — write deferred per Q8 rollback path: C++, Java, Kotlin, plus framework specifics (Spring Boot, Nest, React-strict). When the user invokes the skill on one of these, surface the gap explicitly and propose authoring the reference now or escalating to a follow-up commit.

2. AGENTS.md and per-task GOALS.md split (load-bearing)

• AGENTS.md — project-wide, stable across sessions and tasks. Contains: build/test commands, banned tooling, conventions, contract patterns, and the location pattern of per-task goal files (.agent-tasks/<task-id>/GOALS.md or whatever path the project chooses). Defer authoring the AGENTS.md content to init — this skill only ensures the file exists and contains the goal-location pointer.
• .agent-tasks/<task-id>/GOALS.md — task-ephemeral. Contains: the user's goal in prose, the success criteria the loop checks against, links to the failing tests in .agent-tasks/<task-id>/tests/. Cleaned up after the task merges.

The architectural rule: task A's goals never appear in AGENTS.md. AGENTS.md never contains task-specific success criteria. If the line blurs, surface and refuse.

3. Verifiable-goals scaffolding (task-ephemeral)

For the current task, translate the user's stated goal into failing tests / contract assertions. The tests live in .agent-tasks/<task-id>/tests/ alongside the GOALS.md. The loop runs until the tests pass. Goals therefore exist both as prose (GOALS.md) and as code (tests).

Three sub-shapes, decide while drafting per task:

• Interactive — ask the user what success means, stub failing tests, hand back for review before agent loop begins.
• Template-driven — language-specific test stubs in references/ (per-ecosystem). Pick the matching language template; let the user fill specifics.
• Hybrid — start template-driven; promote to interactive when the user's goal does not fit the template.

The term verifiable goals (Devin Agents101, Jun 2025) is preferred over TDD-for-agents (non-idiomatic in 2026 production stacks).

Cross-references

Pairs with llm-self-loop runtime: this skill runs once at bootstrap, then llm-self-loop runs many times against the gates and goal files this skill installed. The pair is: bootstrap → run-many.

Posture

Bootstrap is a one-shot mode. After running, do not stay resident. The loop that follows is llm-self-loop's territory.